SMF-PR5: Add canonical platform metadata lint gate

.agent-plan.md

@@ -103,10 +103,19 @@ _Source: `docs/external_review/summaries/v1_release_review_synthesis.md` — cro
   - Labels: `type: docs`, `layer: render`
   - Size: S (~200 lines across 4 notebooks)
 
+- [x] **SMF-PR5 / PR 8.4a** — `test(scripts): canonical platform metadata lint gate`
+  - **Add canonical metadata diff/lint step** (HIGH): `scripts/lint_platform_metadata.py` loads the real publication artifacts (`release/kaggle/dataset-metadata.json` + `release/huggingface/README.md`) and fails on Kaggle privacy, license, task, exact platform tag vocabularies, HF split-path/set, Kaggle resource coverage, or task-split schema drift.
+  - **Offline review bundle guard** (HIGH): lint requires root agent-reviewable artifacts plus per-tier `lead_scoring.csv`, `feature_dictionary.csv`, `dataset_card.md`, `metrics.json`, `manifest.json`, task splits, and snapshot-safe relational tables to be represented in the canonical Kaggle resources list. The resource contract now reuses packaging constants for tiers/tables/root docs instead of shadowing the packager.
+  - **File-backed schema check**: when tier CSV/parquet files are materialized, lint compares metadata schema declarations to actual CSV headers and parquet Arrow schemas; `--strict-files` turns missing bundle files into release-readiness failures.
+  - **CI integration**: `release-artifacts-sync` now runs `python scripts/lint_platform_metadata.py` after docs/metrics/claims sync checks, so preview/publish metadata drift blocks PRs.
+  - **Focused tests**: `tests/scripts/test_lint_platform_metadata.py` covers the clean path and injected failures for private Kaggle metadata, license/task/tag drift, HF split paths absent from Kaggle resources, schema mismatch, and missing per-tier review artifacts.
+  - Labels: `type: test`, `layer: validation`, `layer: render`
+  - Size: S (~450 lines, mostly tests)
+
 - [ ] **PR 8.4** — `feat(scripts): integration script + preview hardening`
   - **Regenerate lockfile + bump to v1.0.1** (HIGH): delete `package-lock.json`, update `package.json` pin to `github:ShmuggingFace/ShmuggingFaceCore#v1.0.1`, regenerate via HTTPS. Fixes SSH lockfile and gets the socks/laundry copy fix in one step.
   - **Remove fabricated Kaggle usability scores and medals** (HIGH): delete `TIER_USABILITY`, `TIER_MEDAL` constants from `build_shmuggingface_site.py`. These are dead config today (the framework ignores them) but latent misinformation.
-  - **Make build script read and diff against canonical metadata files** (HIGH): load `release/kaggle/dataset-metadata.json` and `release/huggingface/README.md` in the build script; compare `isPrivate`, tags, license, task, and split counts against the generated config; exit non-zero on mismatch. Structural gap that made `isPrivate: true` invisible in the preview.
+  - **Remaining preview-generator cleanup** (HIGH): `build_shmuggingface_site.py` no longer exists on `main`; if a ShmuggingFaceCore build path is reintroduced, it must consume or validate against the canonical lint gate from SMF-PR5 rather than hard-code task, tags, license, splits, or schema.
   - **Raise on missing/malformed manifest+metrics fields** (MEDIUM): replace `manifest.get("n_leads", 5000)` etc. with explicit key lookups that raise clear errors. For a tool whose job is faithful preview, silent defaults produce plausible-but-false pages.
   - **Use per-tier `dataset_card.md` as each tier's page body** (MEDIUM): currently all three tier pages show the same global README. One-line change per tier in the config builder.
   - **Pin `wrangler` as devDependency + default to preview branch** (MEDIUM): add `wrangler` to `package.json` devDependencies; change `--branch main` to `--branch preview` as the default; add a `--production` flag for intentional production deploys. Prevents clobbering production on every local run.

.github/workflows/ci.yml

@@ -84,6 +84,8 @@ jobs:
         run: python scripts/build_claims_register.py --check
       - name: every claim in claims_register_source.yaml resolves & values match
         run: python scripts/verify_claims_register.py
+      - name: Kaggle and Hugging Face metadata agree with release preview contract
+        run: python scripts/lint_platform_metadata.py
 
   validate-dataset:
     name: Validate lead scoring dataset

scripts/_release_common.py

@@ -363,3 +363,16 @@ def load_relational_column_descriptions(release_dir: Path) -> dict[tuple[str, st
 #: (DGP description, leakage / acceptance bands, break-me guide, etc.).
 #: Copied wholesale into the upload tree when present.
 AGENT_REVIEWABLE_DOCS_DIR: Final[str] = "docs"
+
+#: Required vendored-doc files that should be represented in the
+#: canonical platform metadata. This is separate from the directory
+#: copy contract above: packagers copy the whole docs subtree, while
+#: metadata lint requires these high-value review files to be visible
+#: in the platform file list.
+AGENT_REVIEWABLE_DOC_FILES: Final[tuple[str, ...]] = (
+    "README.md",
+    "feature_dictionary.md",
+    "generation_method.md",
+    "break_me_guide.md",
+    "relational_table_schemas.csv",
+)